MEASURED, not supposed. On 2026-08-03 nine per-app recovery_unit_capture_failed
events reached the hub and TWO operator emails went out. The hub's operator
cooldown key is customerID:eventType(+tier) and that event carries `app` but no
`tier`, so the key held no app identifier: the first refused app took the hour's
slot and every other app's failure was discarded BEFORE anything was written
down, leaving no row on any channel.
The obvious fix — put `app` in the key — was ruled against: on a full disk it
produces one email per app, the volume problem wearing the correctness problem's
clothes.
internal/backup/runsummary.go: a per-run collector with exactly admissionSet's
lifetime, fed by all three write legs, emitting backup_run_failures ONCE at the
end and only when something failed. A clean run emits nothing.
The per-app event stays and becomes the RECORD — the hub routes it record-only,
stored and logged every time, never competing for an email slot. The record and
the notification are now different things.
Deliberate skips (disconnected, decommissioned) are excluded: they have their
own alert, and a nightly email about an unplugged drive is one the operator
learns to ignore.
A manual run always reports: the digest carries a unique run_id the cooldown
cannot collapse. Someone pressing the button is actively trying to get a backup.
THE PERIODIC SWEEP GETS A DIGEST TOO. With the per-app event now record-only, a
capture failure found between runs would be recorded and never notified — a new
silence introduced while closing one. That path emits a digest with NO run_id,
so the ordinary 1-hour cooldown caps it exactly as before while the mail now
lists every failing app instead of whichever was first.
A refusal is recorded ONCE, where the verdict is taken, not at the three legs
that consult it — R-181's contract is one verdict per app per run. Noting it per
leg listed one refused app three times and produced "2 of 1 apps failed". Found
by the digest's own test, not in review.
Silence is safe because the hub's deadline check raises expected_backup_missed
from report freshness, independently of any mail this box sends
(monitor/deadline.go:396,417). Confirmed, not assumed.
7 new tests, 4 red-proofs. The main.go seam walk did NOT fail on its first
attempt — the AST test walked the backup package and not main.go; the test was
fixed and the mutation re-run rather than the pass recorded.
Found by v0.193.0's own live proof run. The estimate was printed fixed to two
decimal GiB, so every app under ~10 MB rendered as "estimated 0.00 GiB write" —
which reads as "no estimate was available" and is the opposite of what happened.
Observed live on demo-hp 08:59:46: opengist's real 178 KB estimate printed as
0.00 GiB.
Shipped in the same session because it is the same defect class R-181 is about:
a message an operator cannot rely on is worse than no message.
The arithmetic is unchanged and still in GiB — the reserve's own unit, so the
comparison against FloorFreeGiB reads directly. Only the rendering moved to
humanizeBytes. estimatedWriteGiB -> estimatedWriteBytes, with the GiB conversion
done once at the point of comparison.
B2's capture floor (v0.192.0) was consulted in exactly ONE place —
captureAllRecoveryUnits, which writes a few KB. The two legs that write the BULK
into the same backups/primary/<app> tree, the DB dump and the volume dump, ran
FIRST and unguarded. Measured live on demo-hp 2026-08-03 06:40:03: opengist's
volume dump wrote 2.0 GB with no check, free fell to 1.0 GB, and the floor then
refused the cheap write it had already lost the argument to. Its refusal message
claimed "the previous unit is untouched" — measured false: that app's tar had
gone 182,272 B -> 2,147,666,432 B under a stale manifest. Sixth entry in
CLAUDE.md's table of shipped guarantees the code did not provide.
Fix: ONE admission verdict per app per run (internal/backup/admission.go), taken
before that app's FIRST write and covering all three legs — they write under one
per-app root, which is why one verdict can honestly cover them.
- Lazy, at the app's first write, NOT once at run start: app A's dump can put app
B under the reserve, so a run-start verdict reads a disk that no longer exists.
- Remembered for the run, never re-decided between an app's own legs — that is
the split this closes. Reset per run.
- Placed ahead of DumpAppVolumesSafe, which stops the stack as its first act, so
a refused app is never bounced. After the volume-less check, which has no write.
- Exactly one operator alert per refused app per run.
- Leg order unchanged: volume dumps still precede the capture.
The floor is now SIZE-AWARE: it asks whether THIS app's write would cross the
reserve, not only whether the filesystem is already below it — which is how an
app was admitted at 96% and then allowed to write 2 GB. Estimate = the app's
previous .sql + .tar on disk. No history -> headroom-only, deliberately, and the
alert says so.
A container-based du per volume was MEASURED and rejected: 66 timed runs on
demo-hp guest 9201, median ~355 ms/volume (341-404) on volumes holding tens of
KB — container start-up, not the walk. Decisive on top: docker run needs the
writable layer, so it can fail under exactly the pressure the reserve handles.
The message was NOT weakened; the behaviour was moved so the wording became true.
It now also names which term bound. Every claim is checked against a sha256
fingerprint of the tree it describes, never against the log line.
Still refuses and never deletes: nothing here is generational.
11 new tests through the production functions. The DB leg cannot run without
Docker, so its gate is pinned by an AST walk of backup.go asserting admitApp
precedes DumpOne (strings.Contains is insufficient — a commented-out call still
contains the string). 4 red-proofs demonstrated failing then restored.
Ships BEFORE the disk-layout merge it exists for, and is harmless on a box
that never gets it. The mp1 partition was a BULKHEAD as well as a ceiling:
it kept a runaway capture from filling the space the container runtime
needs, because /var/lib/docker was a different filesystem. After the merge
it is the same one, and a full Docker data-root is a stopped box.
The floor sits in captureAllRecoveryUnits, checked BEFORE anything is
written: below the reserve, that ONE app's capture is refused, its previous
unit is left byte-identical, the R-158 alert fires with the space figures,
and the loop continues.
Two terms whichever binds first (97% used / 1 GiB free) in fillwatch's
shape, deliberately BEYOND its critical band (95% / 2 GiB) so the customer
is always warned before a refusal can happen — a floor that fires before
its own warning is a silent failure wearing a threshold.
Headroom, never unit size: a per-unit cap would be R-163 rebuilt inside one
volume. Refuses, never deletes: nothing here is generational, so pruning
could only destroy a different app's only local copy; pruneStalePrimaryDirs
is an orphan sweep, not retention, and must not be repurposed.
Tests 1184 -> 1191. One fixture strengthened mid-red-proof: the "old 20 G
ceiling is gone" test sat at exactly 20 GB and survived a literal
UsedGB > 20 cap — hollow. Now 120 GB, and the mutation fails it.
Earned during v0.191.1's own live validation. After the customer had been
warned, a restart produced ZERO fillwatch lines — equally consistent with
'ran and chose silence' and 'never ran'. Proving the checker was alive
needed a deliberate crossing into the critical band.
For an edge-triggered check the quiet run IS the healthy steady state, so
that ambiguity is permanent rather than rare. Check now logs a per-RUN
summary on every run, counting unreadable separately from healthy so a
drive that has quietly gone unreadable cannot read as 'all fine'.
Found live on 9201: neither sched.Daily nor sched.Every fires on
registration, so a box booting with a filesystem already over the line
would stay silent for up to 24h — the R-100 shape, and the same gap the
hub's own checkers avoid by leaving already-breached keys unseeded at init.
The watcher now runs once 90s after startup as well. Safe because the check
is edge-triggered against persisted state: an already-warned filesystem
stays silent. The delay lets mounts settle so a drive still returning reads
as unreadable and is skipped rather than warned about. Pinned by an AST
assertion — the schedule registration alone no longer satisfies the test.
R-167: new internal/fillwatch warns the CUSTOMER before a filesystem fills.
It emits the PRE-EXISTING disk_warning/disk_critical pair, which was
allowlisted, copy'd, default-enabled and checkbox'd with no producer in any
repo — the sixth "built but never wired" instance here. Two threshold terms
(85% or 5 GiB free; critical 95%/2 GiB) because a percentage alone lies at
both ends of this fleet's size range. Edge-triggered on escalation only,
state persisted, hysteresis dead zone at 75%/7 GiB pinned by a test. A nil
usage read is never a warning and never clears one. Per filesystem, never
per app. Daily 03:30, before the nightly app-data legs.
R-158: new unitNotify seam fires per app when a Tier-1 recovery-unit capture
fails, loop continuing, carrying the target filesystem's used/free bytes.
Operator-tier (recovery_unit_capture_failed) — deliberately NOT backup_failed,
which is customer-enabled and would email the customer about a failure they
cannot act on. D-c overrides R-158's own proposal here.
R-174: the app-stop guard no longer starts apps onto MISSING drives — a
regression in v0.189.0 code, found by review and closed the same session.
SetStarter got the raw stack manager, whose StartStack has no drive gate,
and Recover runs at startup. R-171 one path over. bootDriveGate could not be
reused whole (its holder #2 is the guard's own marker, and holders #1/#2 read
vars assigned after Recover runs), so holder #3 is extracted into a shared
driveStartGate with a test pinning the delegation. ErrStartRefused splits a
refusal from a failure: both keep the marker, only Failed alarms, because
routing a deliberate hold into NotifyBackupFailed is the same false alarm.
Tests 1157 -> 1184. All red-proofs demonstrated failing and restored.
Found by live validation on 9201, not by review. GetStacks() is the Manager's
in-memory map refreshed by the scheduler every 10s; sampling it every 5s without
refreshing means two identical samples can mean the cache did not update rather
than that the fleet settled. A container removed ~5s before the window closed was
still in the sampled fleet and the sweep logged 'no boot-orphaned apps' for an app
that had none. sampleBootFleet now refreshes first; a refresh error degrades
rather than aborting the window.
R-171 (a regression v0.189.0 introduced, CONFIRMED on hardware before any fix
was written). Replacing isBootOrphan's container-count term with recorded intent
made a drive-gate-stopped app read as a boot orphan: the gate stops apps with
`compose down` (zero containers) and never touches desired_state, because it is
not the customer. Observed on 9201 with the drive held unmounted — the sweep
found and started it, burned both attempts, and handed it to the dead-app alarm.
The write hazard did not materialise (the unbound mountpoint is host-root-owned
and the guest is unprivileged) but that protection is accidental and untested.
New consumer-side seam bootrecon.StartGate, fail-safe (cannot determine ⇒ do not
start), wired in main.go. The rule is not new: the API's startGatedByMissingDrive
already refuses this; the sweep bypassed it.
R-157 mechanism A. The sweep looked once at T+5s, deriving candidates from a
fleet docker was still restoring — three of six hard resets. Now a settle-then-
sweep window: sample every 5s, settled after 3 identical samples, sweep ONCE at
the end; ends on settled or a 50s budget, and the log says which. The budget is
50s because settle+budget+one retry must stay under the 90s dead-app grace — a
test rejected 60s at 95s. A window that overruns emits a LATE RECOVERY warn
rather than the grace being widened to hide it.
Widening the window made two more holders reachable, so the one gate covers all
three: an absent drive, a quiesce, and an in-flight app-data operation — reusing
quiesce.SuppressedStacks() and a new read-only AppStopGuard.HeldStacks().
R-170. shouldRecreateOnBoot now reads desired_state with the identical three-way
table; absent keeps the old hasContainers behaviour exactly. Its comment argued
for the container count and was rewritten. presentStable is untouched. The two
gates' agreement is pinned from both sides against one fixture table.
27/27 packages green; 6 red-proofs observed FAIL then restored.
The box stops inferring the customer's intent from a container count and reads
what they actually asked for.
Part 1 — desired state. AppConfig gains a tri-state `desired_state`
(""/running/stopped), written ONLY by the customer's own action: the API action
switch, DeployStack, UpdateOptionalConfig's redeploy branch, and the .fab
import. Intent is written BEFORE the act and a failed write REFUSES the act.
StartStack/StopStack are deliberately not writers — 14 callers, only 2 are the
customer. bootrecon.isBootOrphan now reads intent instead of len(Containers)>0,
which closes R-157 mechanism B (a power cut or interrupted deploy left an app
with zero containers, read as a deliberate stop, and stranded silently).
ABSENT MEANS UNKNOWN, NEVER "running": every pre-v0.189.0 app.yaml reads absent,
so the legacy fallback is byte-identical to the old rule. A running-only startup
backfill converges the unambiguous cases; `stopped` is never inferred.
Part 2 — backup.AppStopGuard, a persisted marker over every stop→work→start
window (volume dump, offbox reconstitute, .fab export). Its own file, never
quiesce's. Written before the stop, cleared only after a restart that succeeded,
kept when one fails. Recover() completes before the boot reconciler is launched
and returns its outcome, which main.go reports on the existing backup_failed
event once the notifier exists. A defer is not the mechanism — a SIGKILL runs
none (Campaign 8 fault 10).
Also: SaveAppConfig rebuilt AppConfig field-by-field (the R-100 shape) and would
have dropped desired_state on every save across nine call sites. Replaced with
copy-and-overlay. Measured: app.yaml does not round-trip unknown YAML keys.
No hub change, no agent coupling, no user-visible string. 27/27 packages green;
7 red-proofs observed FAIL then restored.
README's Enforcement block now names controller_gates.py as THE entry point rather than
listing five of the seven gates individually. CHANGELOG records the allowlist entry, the
runner, the hook, the seam test and their red-proofs, and states the no-bump explicitly so
the omission reads as a decision.
A census of all thirteen gate scripts across the four felhom repos on 2026-08-02 found that
every check a CLAUDE.md names was passing and two of the four nobody is told to run were
failing. This repo had seven gates and CLAUDE.md named two; four more were reachable only via
a line in REUSE.md, and docker_run_volume_path_gate.py — RED at census time — through one line
in REUSE.md and nothing else.
controller_gates.py runs all seven plus reuse_refs_check on the repo root, streams each gate's
own output, and exits worst-wins non-zero. The shared reuse checker stays in felhom.eu/scripts/
and is invoked across the workspace — never copied here, which would recreate the drift it
detects; an absent sibling clone FAILS the gate and prints the path tried.
.githooks/pre-push runs it with --fast and refuses the push. Per-clone and --no-verify-able,
both stated in the hook itself; a manual run WARNS when the clone is unarmed.
test_controller_gates.py is a SEAM test — it asserts each member gate's own distinctive stdout,
not the runner's summary line. Red-proofed: an inert run_gate still prints 'all controller
gates OK' and exits 0, and turns the seam test red.
Tooling only: no Go change, no image build, no deploy, no version bump.
realVolumeSize mounts the NAMED VOLUME read-only into a throwaway alpine to du it
from a container view. No host path is involved — docker resolves the volume name
daemon-side — so it is the same safe class as the internal/backup/backup.go entry.
The gate was right to demand review; this diff is that review, on its own, because
burying an allowlist widening inside a feature commit is how an allowlist stops
meaning anything.
Tooling only: no Go change, no build, no deploy, no version bump.
The recovery unit on the customer's drive now carries the PORTABLE secret
class, so Tier-1/Tier-2 restore no longer depends on the whole-guest tier.
A customer needs the drive and nothing else.
Part 0's rulings overturned the brief's recommendation, on evidence:
- the data_key flag is untrustworthy (4+ encryption keys the catalog itself
labels as such are unflagged) -> R-127
- a DB password is not resettable in practice: POSTGRES_PASSWORD is ignored
once PGDATA is non-empty, so a regenerated value leaves the app unable to
authenticate against its own restored rows while the dump replay still
reports success (proven on a throwaway postgres:16-alpine)
Ruling (operator): type:secret travels, type:password never does, minus the
nonPortableSecrets code register. Plaintext -- withholding the internet-
reachable class is what licenses that, and the two are coupled.
Precedence: the UNIT WINS over the guest -- the unit's secrets were captured
in the same run as the dumps beside them, so they match the data being
restored. The fail-closed data-key gate is unchanged.
Secret values are never logged; the manifest records NAMES only.
This is D5's precondition and it is now met.
An app's namespace root IS its backup root: namespaceRoot returns a non-system
drive path as-is, so the recovery unit lands at <HDD_PATH>/backups/primary/<stack>/.
On a NAS that sits inside the share, which FileBrowser binds WHOLE — share root,
:rslave, download:true.
The bind was NOT narrowed, and establishing why inverted the fix. The share-root
:rslave bind is load-bearing (a 2026-07-22 probe proved an in-container access
through it wakes the idle automount trigger), and scoping is undefinable anyway:
apps on a share store at <share>/<app>, there is no userdata/ layer, and creating
one would write Felhom convention onto a customer's own NAS, which R-67 forbids.
So the browsing surface cannot be narrowed and the backup tree must never be
placed under it. Operator ruling: refuse the placement, keep the browse bind.
Tier 2 already refuses network targets for this reason (F-6C-1).
Nothing stranded: zero apps on network storage across all six hub customers
including Peti. R-67's browse capability is byte-identical.
FIVE surfaces, not the four the register named — settings.RefuseAsAppNamespace is
the single predicate. The deploy POST is the real boundary (it accepts any
caller-supplied HDD_PATH; DeployStack validates only os.Stat). Surface 4,
handleStorageDecommission mode=migrate, guarded only its SOURCE, so a whole
namespace could be decommissioned ONTO a NAS — that one is not in the register.
Fails closed: /mnt/felhom-drives holds both kinds, Kind exists only on a
registered path, so an unregistered path under that root refuses.
Supersedes README's "NAS backup locality — decision A" (v0.118.0).
9 tests, all non-effect (nil stackMgr, so a guard that misses panics rather than
passing). 4 red-proofs, each mutation asserted to have landed.
Suite rc=0, 27 packages, 0 FAIL. vet rc=0. Template + emoji gates OK.
Two defects E-2d found on a real box, fixed in this order deliberately: the
message is corrected BEFORE it is put on screen, because switching on a banner
that lies is worse than a silent one.
R-114 — the third state. resolveBackupTargetState had two outcomes: a disk
claims the target (healthy), or nothing does (degraded, "the backup is on the
system disk"). The state "configured, and its drive is gone" had no branch, so
it fell into the second and inherited its message AND its offer. Observed live
with the target detached: degraded:true, target:"felhom-backup" plus the
system-disk copy (false -- the backup was on a drive that had vanished) plus
offer_path naming that same vanished drive as the remedy.
New BackupTargetState.TargetAbsent discriminates. Degraded keeps its meaning
("is there a problem") so the wire contract is unchanged for every consumer;
TargetAbsent answers "which problem", because the two have opposite remedies --
attach any second drive, versus reconnect THAT one. Copy routed through
degradedMessageFor so one place still decides what a customer reads. The offer
is suppressed on the branch itself, NOT left to firstOfferableDrive's
Disconnected skip: that flag is set by the agent-side gate in another repo
(R-113), and this state must be correct independently of it.
R-112 — the state finally has a consumer. The endpoint was byte-correct and
nothing in the product ever asked for it: templates fetch 18 distinct
/api/storage/* endpoints and backup-target[/assign] were the only two with zero
references. Server-rendered on /backups now, following the existing
SingleCopyWarning banner pattern -- not a 19th JS fetch, because a banner that
needs JavaScript to appear is one more thing that can silently not happen.
backupTargetView returns nil for healthy and unknown so those render nothing at
all. The offer control POSTs to the existing assign endpoint behind the standard
inline confirm, never auto-submits, and surfaces restart_required honestly
instead of adding a self-restart.
Scenario E (the seam test) drives backupsHandler over httptest and asserts the
RENDERED HTML -- handler -> view -> resolver -> template. It deliberately does
not call the resolver and assert a string, which would prove the resolver that
was never broken. Deleting the one line that sets data["BackupTarget"]
reproduces the R-112 state and fails every render assertion.
Tests 326 -> 338 (+12) in internal/web; suite green (27 packages); both template
gates pass. Three red-proofs run and reverted, files byte-identical after.
MinAgent unchanged at 0.113.0: R-114 reads BackupTarget/MountPath/GuestPath/Role,
none of which R-113 altered (it changed BoundUnderParent, which this code does
not read). demo-hp on agent 0.113.0 is not held.
The absent copy is verbatim the hub's customerMessages["backup_target_absent"]
so the banner and the email tell one story -- filed as a two-repo drift risk,
not solved.
NOT LIVE-VALIDATED. Scenario C cannot occur on a healthy box; Session C proves it.
Registered as /api/backup-target inside ServeStorageAPI, which main.go mounts ONLY
at /api/storage/. Live result: endpoint not found, while every unit test passed --
the tests called the handlers directly and never travelled the mount. Caught by
the first live call, which is why the live call is part of the procedure.
Moved to /api/storage/backup-target[/assign]. A new test asserts the dispatcher
source contains both paths, so a handler nothing routes to fails the suite --
the seam-wiring rule applied to a route rather than a button.
The half that makes the rest work: a degraded backup target recorded only in
config is the silent-degradation pattern this arc has spent a week removing.
Part 3 -- POST /api/backup-target/assign moves the target via the agent's
POST /backup/target. It is the ONLY writer of the role: registration does not set
it, the drive-gate does not, no scheduler does. Declining is not calling it. The
agent returns restart_required rather than restarting itself, because restarting
with a backup in flight records a spurious tier failure for a backup that
actually succeeded (E-1 did exactly that).
Part 4 -- GET /api/backup-target returns the state and, when degraded, Hungarian
copy in FACT -> CONSEQUENCE -> REMEDY order, pinned by a test: a customer told
only the fact cannot act on it.
Healthy renders NOTHING -- no badge, no reassurance, no tonal change.
degradedMessageFor is the single decision point, so exactly one place could start
decorating a working box. Red-proofed: reassuring on the healthy branch fails
Scenario E.
UNKNOWN is not degraded: an unreachable or pre-R-82 agent means we could not ask,
which is not evidence of degradation (R-88 Part 2's class).
A HOLLOW TEST caught by its own red-proof: TestUnknownStateRendersNothing used
{Known:false} with Degraded left false, so it passed even with the !Known guard
deleted -- the second condition covered for it. Now {Known:false, Degraded:true},
which fails properly. Without the red-proof the test would have been decoration.
State is derived from the AGENT, never from our intent flag: on the two boxes
migrated by hand in E-1 the intent was never recorded while the drive really is
the target.
MinAgent: 0.113.0
Green gate: build + vet + test rc=0 (27 packages), run separately from this commit.
Caught before deploy by tracing a.Path back to its source, not by a failure. The
0.184.0 image is superseded and must not be shipped.
ReconcileDriveGates resolves the target as isTarget[a.Path], and a.Path is the
REGISTERED StoragePath -- for an external drive that is the GUEST path
/mnt/felhom-drives/<name>, not the agent's host MountPath (/mnt/<name>) that
/disks reports. driveTargetByPath keyed on MountPath alone, so the lookup never
matched: every absent drive, the target included, fell through to the generic
storage_disconnected.
The alarm would have looked wired, passed its own unit tests, shipped, and been
silently wrong on exactly the drive it exists for -- the same defect class E-2b
was opened to fix, one level down.
Now keyed under BOTH paths, mirroring planDriveGates which already registers
present[] under GuestPath and MountPath for the same reason.
Red-proof: MountPath-only keying fails with "the backup target is not resolvable
by its GUEST path -- the gate passes a.Path (the registered StoragePath), so the
backup-target branch would never fire".
Green gate: build + vet + test rc=0, run separately from this commit.
NotifyStorageDisconnected and NotifyStorageReconnected were defined and called
from NOWHERE. Registered in allowedEventTypes, in DefaultEnabledEvents, and given
a Hungarian message on the hub -- and never invoked. A drive going absent produced
apps stopped, a WARN log and a UI badge, then silence on every channel. Verified
against the gitignored-cmd/ trap with a positive control. Fifth instance of this
class, found by E-2 Phase 0 rather than by a failure.
A drive that is ONLY a backup target has no apps to stop, so it was silent twice.
ReconcileDriveGates now calls both halves. When the absent drive is the whole-guest
backup target it raises the more specific backup_target_absent (error) instead --
never both, since two mails for one event trains people to ignore the channel --
and recovers as backup_target_restored (info, the existing pairing-gated pattern;
severityNotifies NOT widened). The recovery mirrors the alarm's choice or the
operator cannot match them.
Which drive is the target comes from the AGENT (/disks backup_target, >= 0.112.0),
not from our StoragePath.BackupTarget: that is customer INTENT, and on the two
boxes migrated by hand in E-1 the intent was never recorded while the drive really
is the target. An older agent omits the field -> false -> generic alarm, never a
wrong one.
Before this an absent backup target had NO prompt signal: the tier stays DUE
(targetStoragePresent checks name presence, never reachability), so the only
evidence was its own failure at the next due cycle, up to ~24h away. The R-100
shape.
Tests observe the WIRE, not a mock, because the failure class is "nothing
arrives": a real Notifier posts to an httptest hub and the test asserts the event
type and severity that actually went out.
MinAgent: 0.112.0
Green gate: build + vet + test rc=0 (27 packages), run separately from this commit.
Foundation only, no version bump -- nothing customer-visible changes yet. The
offer UI (Part 3), the degraded banner (Part 4) and the controller half of the
absent-target signal (Part 5) are NOT here; they are tracked as E-2 in
OPEN-ITEMS.md so this cannot become a sixth "seam built but never wired". The
fifth was found by E-2's own Phase 0: NotifyStorageDisconnected is defined and
called nowhere, so a drive going absent emits no event at all today.
StoragePath gains BackupTarget bool -- the sibling role to Schedulable/
IsDefault/Kind. It is INTENT, not truth: the authority is the agent's
backup.local_backup_target, and truth is read from GET /backup/tiers. This
records what the customer ASSIGNED so the controller can render the state,
notice the drive going absent, and detect drift.
Invariants, each pinned by a test asserting the CONSEQUENCE not the mechanism:
- a drive NEVER acquires the role by appearing (registration does not set it)
- exactly one carrier; assigning moves rather than duplicates
- sticky: a new bigger/faster drive does not steal an assigned target
- an absent target STAYS assigned -- clearing on disconnect would be a silent
retarget by omission ("no target configured" instead of "drive missing")
- a network share is refused (R-108 risk model; the role is about LOCAL disk
failure)
Red-proof C: adding auto-elevation to AddStoragePath fails
TestRegisteringDrivesNeverAssignsTheBackupTarget with
registering drives assigned the backup target "/mnt/hdd_1"
Attributes may suggest and refuse the absurd, never select: demo-felhom's backup
drive is an external USB HDD and BOTH demo boxes report removable=0, so a
transport rule disqualifies the reference drive and a removable rule finds no
candidate at all.
Green gate: build + vet + test rc=0, run separately from this commit.
Both are the system reporting healthy while the customer is not, and both live in the same
status-derivation code. Neither is fixed by making the system quieter.
C9-F1 (HIGH) — Tier-2 writes recovery-unit/ on EVERY run and RestoreTier2Files has never read
it (tier2_restore.go:101-104 reads hdd/ + userdata/ only). Phase 0 enumerated all 53 catalog
templates against both demo boxes: 43 apps have NO readable subtree, so the button stopped the
app, restored 0 files, restarted it and said "Nincs hiányzó fájl — minden fájl megvan a helyén."
— at the moment the customer pressed it because files were missing, with 156 MB of BookStack's
data unread in the same copy. 9 apps have file legs but never their DB or volumes, so the same
sentence was also a clean bill of health over data never opened (immich: 1.3 GB Postgres unit).
Honesty half shipped: a pre-flight coverage check refuses UP FRONT without stopping the app and
NAMES the action that works; a run that proceeds claims only what it EXAMINED and discloses that
the database and volumes are not covered. Completeness is filed as C9-F1b — routing to the
Tier-1 unit restore puts a destructive operation behind a non-destructive button, so its confirm
copy has to carry that difference. C9-F4 filed: nothing reads the Tier-2 recovery-unit/ mirror,
so the second local copy that exists for drive loss is unreachable by any customer action.
C9-F2 (HIGH) — a crash loop was counted as working. StateRestarting is deliberately NOT added to
IsDownState (that alarms on every deploy fleet-wide, the over-correction F-A1 nearly cost us); a
sustained run becomes down after crashLoopAfter = 5m, set above the 120s deploy timeout, Mealie's
60s start_period and R-97b's 180s grace. The dashboard counter uses the same predicate, so it no
longer contradicts the alarm on the same screen. README's claim that faults "still surface as
restarting" was a wish with no test — corrected in place; it is the seventh such instance.
Six red-proofs observed, including the one that matters most: adding StateRestarting to
IsDownState fails the brief-restart test with "every deploy and update would page the operator".
go test ./... rc=0, 27 packages, run and read separately from this commit.
Tier2LastRun is the attempt clock and was rendered as 'Legutóbbi másolat' in the restore
confirm dialog. New LastSuccess + SuccessTracked anchor; tier2Update makes the three
rebuild sites safe by construction. F-DIAG: six distinct causes, target-aware redaction.
LastRun records an attempt, not a result. New OffboxTarget.LastSuccess, set only on the
success branch via the pure offboxAnchorAfterRun rule, carried to the hub as last_success.
Closes two silent-wipe sites (settings save, hub re-apply).
deadapp-check had no observable at default info level: its per-cycle line goes
through Scheduler.dbg(), gated on logging.level==debug, so on a default box it is
never PRODUCED (not merely filtered) and cannot reach the always-DEBUG ring. A
30s interval also puts it on the scheduler's quiet path. 'No alarms' was
therefore indistinguishable from 'the detector never ran' — which undermines
confidence in the F-CRIT-1 fix in the field.
A periodic summary, not a line per run: at 30s a per-run line is 2880 lines/day,
which is why the original author chose silence. Every 20th scan (~10 min) emits
one INFO with the scan count, apps evaluated and apps down. A test pins the
cadence so it cannot be widened into uselessness.
Also corrects the 'unquiesce guaranteed by defer' comment — fault 10 established
the guarantee is the crash marker plus Recover().
F-CRIT-1 — an app that failed to restart after a quiesce never alarmed, for two
independent reasons, either of which alone kept it dead:
1. restartAll returned nothing, so the failure was logged and dropped and no
caller could learn a customer's app had not come back. It now returns the
stacks that failed; both call sites record the outcome.
2. classifyRunStates whitelists StateStopped on invariant I1 ('StateStopped
means the user stopped it'). The quiesce loop stops stacks by the same
compose-down path, so a failed restart is also StateStopped and was
whitelisted into silence. Loop.FailedRestarts() is now the only thing that
lifts the whitelist, so genuine user stops stay silent (v0.164.0 pinned).
F-A1 — HTTP 409 is the agent's single-flight gate refusing while a restore-test
holds it, not a failure. agentapi now returns a typed *StatusError on POST, the
adapter maps 409 -> quiesce.ErrTierBusy, and the loop defers: no breaker, no
event, no operator email, tier stays DUE.
Two traps avoided. Silence: contention outliving contentionAlarmAfter (3h, set
by the agent's own 120m PBS restore-test ceiling) raises its own BLOCKED signal.
App thrash: removing the failure treatment also removes the breaker's deferral,
so a contended tier is dropped BEFORE anything stops (contentionRetryAfter 15m,
against a 12m01s longest observed restore-test).
Three comments corrected; the invariant rule added to both CLAUDE.md copies.
Six red-proofs, all observed failing.
A witness in a _test.go file fires on go test and go vet but NOT on go build
alone — and a build-only step is exactly how the R-88 Part 2 near-miss would have
shipped. Moved beside the type it pins, and added one for AgentVersionReporter.
No defect found: quiesceBackend and *Client both satisfy their interfaces today.
No version bump, no deploy — compile-time only.
MinAgent: 0.105.0. scheduledRunAllowed fired on any nil age; it now requires a
licence from valveLicensed, which grants it for AgeStateAbsent and for a LEGACY
agent, and refuses it for AgeStateUnknown. An unreadable storage no longer
masquerades as a first-ever backup and no longer quiesces apps outside the window.
A missing wire field means legacy, not unknown — deliberately. Treating it as
unknown would stop the valve firing on un-upgraded boxes and starve genuinely new
ones. Degrade logged once; unrecognised future values also map to legacy.
Caught in passing: TieredBackend is satisfied by a RUNTIME assertion, so the
signature change compiled and vetted clean while quiesceBackend silently stopped
satisfying it — which would have degraded every box to the single-tier path with
no error. Added a compile-time witness.
Also corrects the notifier comment that claimed operator-only came from a missing
customerMessages entry; enforcement is hub-side operatorOnlyEvents (hub 0.79.0).
R-97a: internal/quiesce had no route to the hub at all — three failed whole-guest
backups on 2026-07-27 produced zero events. TierNotifier is a seam (not an import),
wired by an init-only setter because main.go builds the notifier after the loop.
Edge-triggered: the failure fires when the R-88 breaker ARMS, not per retry, and
recovery rides recordSuccess's existing bool. Uses NEW operator-only event types;
reusing backup_failed would have emailed the customer in Hungarian about a backup
they cannot act on, since it has a customerMessages entry and is in live
enabled_events. Requires hub >= v0.78.0.
R-97b: v0.164.0's state filter cannot see an app caught MID-RESTART, which is how
BookStack alarmed. The fix is a suppression window keyed to the quiesce CYCLE,
consumed at the same single derivation point. 180s grace, derived from the deploy
flow's 120s health timeout and Mealie's 60s start_period; it expires, so an app
that genuinely fails to come back still alarms.
internal/quiesce had no failure counter, no backoff and no breaker, and the driver
is a plain 5-minute ticker — so a tier that was due and kept failing stopped and
restarted every customer app stack every 5 minutes indefinitely. Live on
demo-felhom 2026-07-27: three cycles in eleven minutes against an unreachable PBS
tier; it ended only because PBS recovered.
The breaker gates the QUIESCE, not the backup — the harm was the outage taken to
attempt it, so backed-off tiers are dropped before any stack is stopped. Per
target (a broken offsite tier must not suppress a healthy local one), 15m→30m→
1h→2h→4h capped, reset on success, never permanent, never applied to TriggerNow,
and stillRunning is not a failure. State is in-memory on purpose: forgetting a
backoff costs one attempt; persisting one could outlive the fix.
Part 3 invariant recorded on scheduledRunAllowed — a missing value means UNKNOWN,
and only a positive 'never' may fire the safety valve. Fourth instance of the
class (hub v0.12.0, v0.73.0, R-81, R-88).
Part 2 (unknown != never) is NOT in this commit: the agent returns byte-identical
responses for 'read errored' and 'never backed up', so the controller cannot tell
them apart. That needs an agent wire change and is tracked separately.
Operator ruling 2026-07-26: let the first backup run as long as needed; other
backups shouldn't start until finished.
A first FULL offsite snapshot runs for hours, far past max_quiesce. When that
bound elapses the app resumes (unchanged), but the loop then started the NEXT
tier while the first was still uploading. Now it breaks and defers the rest to
a later poll — vzdump still holds the guest lock, so the second start would be
refused by the agent (409, v0.99.0) or fail on the lock, and a failed backup
never satisfies a cadence, so the tier would retry into the same wall forever.
pollTier returns (phase, stillRunning, err). The app still resumes exactly once.
Red-proof observed and restored; full suite green.
MinAgent UNCHANGED — degrades gracefully against ANY older agent.
The agent gained per-target tiers in v0.97.0. The controller owns quiescing,
so the multi-tier schedule is reconciled here: every due tier is collected up
front and run inside ONE quiesce window (one stop, N sequential backups, one
resume). Two cycles on the weekly night would mean two app outages for one
night's work.
Dedup rule: local-only -> one quiesce; PBS-only -> one quiesce; BOTH due ->
ONE window with both backups inside; neither -> no quiesce.
- quiesce.TieredBackend + BackupTier + ErrTiersUnsupported (optional extension)
- agentapi: BackupTiers/BackupDueFor/StartBackupFor/BackupStatusFor;
targetQuery("") yields an EMPTY suffix so untargeted hits the pre-R-82 route
byte-for-byte
- Loop.resolveDueTiers = the dedup rule in one place, agent order preserved
- quiesceAndPollTiers + pollTier: app stays quiesced until the LAST tier
snapshots (resuming earlier loses app-consistency on the DR tier). Consequence
stated in the docs: both-due-night downtime = first tier's full backup + last
tier's snapshot, which is why tiers run fast-first.
- Manual 'Mentes most' covers EVERY tier, due-ness ignored.
- Window-gate safety valve now uses the OLDEST due tier, so a stale DR tier
cannot be starved by a fresher local one.
Capability detection: /backup/tiers 404 = pre-R-82 agent (the documented
route-probe mechanism). Not a featureProbes row on purpose — the loop needs the
tier LIST, not a yes/no. Degrade logged exactly once per process.
Tests +11, full suite green. Red-proofs #2 and #3 observed and restored.
Source: felhom.eu/documentation/audits/DIAG-agent-channel-2026-07-26.md
bootstrap.DetectEndpointDrift names a controller.yaml vs bootstrap.json
local_api.endpoint divergence -- one ERROR carrying BOTH values and BOTH paths,
its own event type local_api_endpoint_drift, and its own Hungarian banner shown
ABOVE the channel banner because drift is the cause and "agent unreachable" the
symptom. It writes NOTHING: reconciling from bootstrap.json would clobber a
correct controller.yaml on any half-provisioned or hand-repaired guest, so the
authority ruling is deferred to R-78. Fail-safe silent on absent/unparseable/
incomplete bootstrap and on an empty endpoint (ensureLocalAPI's fill-if-missing
path is untouched). Fingerprint compared as a BOOLEAN only; token never
compared, logged or exposed.
EffectiveProtected now gates samba on Enabled && UserSet, mirroring BOTH of
reconcileSambaAt's early returns, and the doc comment is corrected in the same
change -- it claimed "detection and deployment agree in both directions" while
citing only !smb.Enabled, an assertion that went false when !smb.UserSet was
added. Not over-suppressed: sharing on WITH a password and a dead container
still alarms.
Channel log: the debounce placeholder is stateUnconfirmed (rendered "unseeded")
instead of "up", so a born-down channel no longer logs "up->down" and orUnseeded
stops being dead code. Logging only -- the placeholder is still matched in the
re-arm condition, so F2 born-down alerting is byte-for-byte unchanged and all
nine pre-existing channelhealth tests pass.
Tests 951 -> 959, all green. Red-proofs A (both directions), E and F.
MinAgent unchanged; felhom-agent untouched.
Found on the demo-hp live leg: with import, import/paperless and import/calibre
in the carry-list, the derived skeleton RE-CREATES a per-drive drop-zone on every
drive forever — the dead lookalike the canonical root exists to remove, and one
that is never backed up (class: excluded).
Not a zero-removals violation: nothing deletes what an existing box has. Both
demo boxes' old drop-zones were verified to hold zero files before the change.
Found on the demo-felhom deploy leg: ensuring only <sysNS>/userdata/import left
its parent at 755 root:root, because EnsureUserdataDir MkdirAll's intermediates
at plain 0755 and chmods only the leaf. That made the system drive's userdata
root the one on the box outside the 2775/gid-1000 convention.
${IMPORT_PATH} = <system namespace root>/userdata/import — ONE drop-zone per box,
on the system drive, injected at BOTH compose-env builders with NO per-drive
fallback (unresolvable leaves it unset so compose fails loudly rather than
quietly building a second, dead drop-zone).
Third BindRoot (RootImport) + Import list in BackupSpec, extended through
ValidateBackupSpec/ClassifyBinds. Load-bearing: a stale `userdata: import/<app>`
entry against the moved bind would be a WHOLE-BLOCK reject, taking the app's
mandatory hdd classification with it.
Exhaustive-root audit: resolveAbs/structuralGuard/ComputeCaptureSet/
ComputeFabBuckets now take importRoot explicitly (an import bind resolved
against hddPath would name a directory on the wrong drive); unresolvable is
refused loudly into Skipped. GetImportRoot added to both provider interfaces.
Catalog-derived skeleton: UserdataSkeleton() -> UserdataSkeletonCarry() +
BuildUserdataSkeleton(), SORTED. The carry-list makes zero-removals true by
construction (`documents` is in no catalog app but on both boxes) and is the
fresh-box floor. The sort is not tidiness: the naive map-order derivation
measured 20 distinct outputs from 20 identical runs, which with fbNeedsRecreate
is a fleet-wide FileBrowser restart loop.
One authoritative compose parser: ParseComposeUserdataMounts now delegates to
ParseComposeClassifiableBinds. Import root excluded from per-app migration.
Surfaces: FileBrowser /srv/beolvasas source; app-page "Hova tegyem a fajlokat?"
with PathEscape deep links (never QueryEscape) and class-driven copy;
data_paths: annotation with the Fork-3 asymmetry; system-owned beolvasas SMB
share refused server-side at handler AND store, button omitted in template.
Caught on the way: the sharing template's row struct was function-local, so
adding {{if .System}} would have 500'd every share row. ShareRow is now
package-level and the render test uses the handler's own type.
Tests 915 -> 949, all green. MinAgent unchanged.
agentapi.SmartSummary.ModelName mirrors the agent's model_name; the card row label
prefers the device model over the raw name/UUID, falling back to Name(+hint) on an
old agent. Additive. Test + red-proof (drop fallback -> A4 fails).
/ now 302s to /launcher (ONE canonical URL per page — launcher body never served at /);
the Vezérlőpult keeps /dashboard + its nav slot (operator ruling, reversing v0.163.0).
Nav: Indítópult active on /launcher, Vezérlőpult href=/dashboard active there. Post-login
default (/) and the topbar logo (/) flow through the redirect; login target unchanged.
Repointed 2 dashboard-card tests /→/dashboard. Corrected the stale vacation/agent-DOWN
note in CLAUDE.md (agent is up at 192.168.0.162:8443).
Tests: 302 target+status, /dashboard 200, nav hrefs/active; red-proof verified.
gofmt -w across the controller tree (46 files) so gofmt -l is empty — disarms the
formatting landmine where a targeted edit + accidental gofmt -w swept ~46 unrelated
files. Pure formatting: whitespace + gofmt's optional-semicolon removal in reflowed
inline closures. One doc comment reworded ('' -> 'the empty string') to avoid gofmt's
Go-1.19 doc-comment typographic substitition ('' -> curly quote) muddying its meaning.
No build/vet/test behavior change.
The agent defaults SMART to UNKNOWN on non-physical targets, so they showed as
spurious 'Nincs adat' rows. isPhysicalDisk now excludes those types (card + check).
Test strengthened: a PBS/LVM fixture with UNKNOWN SMART must still be excluded.
Consumes the agent v0.94.0 smart payload (MinAgent floor unchanged; feature-detect
by presence). One pure verdict fn agentapi.DiskVerdictFor shared by the dashboard
card and the 6h check. Card via a 60s /disks TTL cache (anti-smartctl-storm);
unreachable agent -> Nincs adat, page never blocks. disk-health-check (6h) emits
disk_health_degraded on a degradation only vs an in-memory baseline (first run
silent, recovery/UNKNOWN never notify, multi-attr -> one event). No global banner
(deliberate). Pairs with the hub allowlist bump.
Tests: verdict table (>=90 red-proof), notifier emit, check first-run-silent
(red-proof), degradation-once, recovery-silent, UNKNOWN-excluded, FAILING-critical,
nil-smart card, TTL cache.
ONE setting (window start W) drives every nightly leg at fixed, never-stored
offsets: DB dump at W, tier-2 at W+60m, off-box at W+105m (wrap-safe). Precedence
settings > controller.yaml db_dump_schedule > 02:30.
- scheduler.UpdateDaily: retime a daily job at runtime (no restart) via a per-job
buffered resched chan + a select case in runDailyJob.
- new pure package internal/backupwindow (LegTimes/GateWindow/EffectiveWindow).
- quiesce disk-tier window gate: scheduled cycles run only inside [W+2h,W+6h) with a
safety valve (age>cadence+24h runs regardless); manual TriggerNow never gated.
Backend.Due now also returns the backup age (from the agent's own /backup/due).
- backup page: Mentési időablak card (time input + derived leg/gate rows); POST
/backups/window validates -> saves -> UpdateDaily x3 -> flash.
Tests: 5 groups, all red-proofed. Agent/cadence//backup/due untouched.
.sidebar-logo was left-aligned; add auto side-margins so the 140px logo centers in
the header (desktop sidebar + mobile drawer share the element). Adds TestSidebarLogo_Centered.
Viktor pushed the text-outlined logo.svg to felhom.eu main (be9edb4) — wordmark is
now 17 real path glyphs. Swapped FelhomLogoSVG to it; stripped Inkscape's leftover
empty <text/> shells + font-* style leftovers on the paths via lxml DOM pass (glyphs
untouched, no text-to-path by CC), dropped editor namedview. Cleaned FelhomFaviconSVG
vestigial <text>. Both constants now free of <text>/font-family; viewBoxes unchanged;
palette + gradients preserved. TestLogoSVG_NoLiveText committed green.
Mobile nav was broken — the <=768px block predated the v0.146.0 accordion and
flattened .nav-links into a horizontal overflow-x strip, clipping the accordion's
nested sub-lists. Replaced with a sticky top bar + off-canvas left drawer that
reuses the vertical sidebar (accordion untouched), plus a no-js static fallback.
Removed the sidebar customer-name span (kept on login). Added ?v={{.Version}}
cache-bust to logo/favicon URLs (Cloudflare 4h edge-cache; 0.126.1 failure mode).
Part 4 (outlined-logo constant swap) gated out per §3a: live felhom.eu main still
serves a logo.svg with live <text>/font-family; constants unchanged.
5 new tests via the real layout/CSS render; nav_accordion invariants unchanged.
Feature-detected navigator.share opens the OS share sheet with title+text+URL only
(no QR files: — narrow support drops the URL when given file+URL). Hidden unless
supported; Link másolása stays the universal fallback and catches the non-cancel
rejection; AbortError is silent. Template JS + tests only. 2 red-proofs verified red.
Mint a 160-bit capability URL (/s/<token>) serving a standalone read-only guest
launcher: same tiles, opens apps in new tabs, no account, no admin session.
Information only, zero control — every privilege stays behind each app's own auth.
- /s/ pre-auth pass-through (after the claim gate) + session-CSRF exemption; guest
password POST carries its own pre-auth HMAC CSRF.
- Constant-time token match; empty stored token = disabled = byte-identical mux 404.
- Optional per-share password: separate bcrypt hash + own attempt map; signed cookie
= HMAC(token|passwordHash) keyed with web.session_secret, so rotate/change invalidates.
- Guest labels ride the v0.164.0 ruling; never expose internal state vocabulary.
- Token redacted in logs (/s/<redacted>); never in CHANGELOG/REPORT/CONTEXT.
- Admin modal: copy-link, QR (go-qrcode), set/clear password, rotate, disable.
- Tests: Groups A-G (14) + 3 red-proofs verified red.
A UI stop (Leallitas -> compose down -> StateStopped) is the user's own
action, not a fault, and must not raise the deadapp banner OR the
app_start_failed event. Filter at the single fix-3 derivation point:
extract scanDeployedAppRunStates's pure core to classifyRunStates and
change the down predicate to IsDownState(st.State) && st.State !=
StateStopped. Suppresses StateStopped from both the banner dead-list and
the notifier Down-set at once.
Rests on two invariants (recorded at the seam, README, CONTEXT):
I1 StopStack = compose down => zero containers => StateStopped
I2 P2 census: all catalog services unless-stopped => faults never rest
at stopped (they surface as exited/degraded).
IsDownState unchanged; out-of-band 'compose stop' (containers remain ->
exited) still alerts. Tests +4 (notify 3->4, main 4->7), both red-proofs
verified. No template/funcmap/notifier/counter/copy change.
(A) The launcher monogram bled through every transparent white logo — it rendered
unconditionally under the img. Now hidden by default (.launch-mono display:none),
revealed only when the img chain fails (onerror adds .launch-tile--noimg).
(B) The /static/app-placeholder.svg default reached only app_list_row. The four
other sanctioned app-logo onerror chains now match the canonical SVG->PNG->placeholder
grammar: backups_apps (aligned row), stacks (infra->infra-logo else placeholder),
app_info hero (screenshots still vanish), deploy (keeps .LogoURL/.LogoPNGURL).
Template/CSS only; no handler/funcmap change. 5 tests + 2 red-proofs.
New /launcher page: a grid of large tappable tiles, one per openable deployed
app (subdomain presence is the single openability criterion, shared with the
Megnyitás button via the extracted Server.subdomainMap helper). Colored tiles
(deterministic slug color or .felhom.yml brand_color), white glyph/monogram
fallback, target=_blank links for operational apps, greyed unclickable tiles for
stopped ones. First sidebar item; / stays the Vezérlőpult.
Universal app placeholder: new AppPlaceholderSVG served at
/static/app-placeholder.svg, now the default FallbackIcon on app_list_row so a
logo-less app shows a placeholder instead of visibility:hidden. Brand mark is
never an app placeholder.
New Metadata.BrandColor; new funcmap tileColor/initial. 10 new test functions +
4 red-proofs. No agent coupling; MinAgent unchanged.
admin
Claude Code